Jais-3-32B

Jais-3-32B is the large dense member of the Jais-3 family: a 32B-class decoder-only model with a 512K context window.

Other formats are available in the IFM/Jais-3 collection:

• Quantization: IFM/Jais-3-32B-FP8
• GGUF for llama.cpp and Ollama: IFM/Jais-3-32B-GGUF
• MLX for Apple silicon:
  • IFM/Jais-3-32B-MLX-4bit
  • IFM/Jais-3-32B-MLX-8bit
  • IFM/Jais-3-32B-MLX-bf16

Benchmark Results

		Reference models
	Jais-3-32B	Falcon-H1-34B-Instruct	Gemma-4-31B-it	Muse-Glimmer-30B	Qwen3.8-27B	Fanar-2-27B
# Params	32B	34B	31B	30B	27B	27B
# Activated params	32B	34B	31B	30B	27B	27B
Architecture	Dense	Dense	Dense	Dense	Dense	Dense
Open-ended Generation
AraGen
AraGen v3 (3C3H) 09-2025	71.0	45.0	73.4	75.6	66.7	55.9
Correctness	81.0	53.2	85.4	87.7	77.1	65.7
Completeness	80.3	51.0	84.9	87.0	76.2	63.5
Conciseness	28.1	19.7	21.1	23.0	23.9	21.6
Helpfulness	76.4	45.4	79.7	81.2	71.5	59.0
Honesty	79.9	49.0	84.4	86.7	74.7	61.3
Harmlessness	80.3	51.4	84.9	87.7	76.5	64.4
Knowledge & Reasoning
General Knowledge & Comprehension
Arabic MMLU* Accuracy	88.6	73.7	89.5	87.0	85.8	66.3
Belebele Accuracy	95.0	86.6	95.9	94.3	94.4	61.8
STEM & Math
AraSTEM* Accuracy	95.2	85.5	96.3	94.7	95.3	42.4
Arabic GSM8K* Accuracy	94.3	90.2	93.7	95.2	93.2	87.7
GATmath* Accuracy	98.5	63.4	98.6	98.5	98.4	87.8
Professional Domains
MedAraBench* Accuracy	87.0	60.7	88.8	86.1	87.3	50.9
Arabic Legal (HELM)* Accuracy	73.4	63.1	71.1	73.1	70.5	45.7
Arabic Language & Culture
Islamic Knowledge & Arab Culture
IslamicQA Accuracy (mean of the 4 below)	88.7	71.1	89.7	83.7	77.1	52.0
PaLM-X 2025-2* Accuracy	94.6	84.7	94.5	92.6	90.8	69.3
QIAS-1* Accuracy	95.9	56.2	93.2	87.4	81.6	31.4
QIAS-2* Accuracy	85.5	73.4	86.9	83.0	81.9	61.8
IslamicEval2025-1B* Accuracy	78.8	70.1	84.0	71.9	54.1	45.5
Arab Culture Accuracy (mean of the 3 below)	88.0	76.5	90.2	84.6	81.0	41.4
ArabCulture* Accuracy	88.2	79.7	91.4	86.0	85.0	34.2
AraDiCE-Culture* Accuracy	88.3	75.4	89.4	83.2	77.1	55.9
Arabic Cuisine^ Accuracy	87.4	74.4	89.8	84.7	81.0	34.0
Dialects
Dialectal Arabic MMLU* Accuracy	88.5	71.5	89.4	87.0	87.0	28.2
Dialect Identification (QADI) Accuracy	70.6	50.3	70.4	60.6	51.5	29.6
Alyah* Accuracy	83.0	71.8	88.4	80.3	77.7	36.3
Linguistics & Literature
Arabic MMLU with diacritics Accuracy	87.9	69.1	88.9	86.6	85.2	64.2
FannOrFlop-MCQ Accuracy	71.8	62.4	72.6	68.3	62.5	46.4
Arabic Linguistics^ Accuracy (mean of the 2 below)	94.3	57.4	96.7	91.8	89.7	39.1
Diacritization Accuracy	91.9	65.2	95.2	89.5	89.6	47.0
Parsing Accuracy	96.7	49.7	98.3	94.1	89.8	31.3
Translation
FLORES200+ LLM-as-a-judge	91.5	70.6	92.8	84.8	81.8	71.2
Assistant Capabilities
Instruction Following
Arabic IFBench^ Accuracy (prompt-level strict)	88.7	32.2	80.8	84.2	81.3	40.0
Arabic IFEval Accuracy	80.0	66.8	78.7	79.5	77.0	49.8
Long Context
AA-LCR (Ar) Accuracy	75.0	14.0	66.0	-	79.0	-
Factuality & Hallucination
HalluScore Accuracy	67.3	42.4	69.7	69.1	54.6	41.7
SimpleQA Verified (Ar) Aggregate (2/3 F-score + 1/3 non-hallucination rate)	28.1	7.5	10.6	13.3	8.0	11.0
F-score	7.1	6.6	10.7	18.2	9.2	10.4
Non-hallucination rate	70.1	9.4	10.5	3.7	5.4	12.2
Agentic
Terminal Bench 2.1 Accuracy (Avg@3)	63.7	1.1	43.5	51.7	79.8	1.1
GDPVal Accuracy (@ 75% threshold)	75.5	0.0	30.5	48.2	78.2	0.0
MCPMark Accuracy (Pass@1)	66.1	0.8	48.8	44.9	49.6	12.1
WildClawBench (English text-only subset) Mean Score@3	42.0	10.7	44.1	43.2	52.1	10.3
Apex-Agents (420 text-only subset, loop agent) Pass@1 (Accuracy)	25.2	0.7	12.4	22.1	28.6	0.2
GAIA v2 lit (Ar) Accuracy (Pass@1)	51.5	9.7	42.4	36.4	55.8	12.1
Arabic File Processing Accuracy (@ 75% threshold)	71.8	0.0	4.5	20.9	78.6	0.0
SWE-Bench Verified (Offline) Accuracy (Pass@1)	72.4	0.4	59.0	71.4	73.5	0.4
BrowseComp-Plus Accuracy	71.9	1.8	42.5	41.1	53.0	3.9

Starred (*) benchmarks have been audited for sample quality. Items with ambiguous questions, invalid answer options, or indefensible gold labels were excluded, since they can penalize a correct answer or reward a lucky guess. Scores on these benchmarks are reported on the cleaned sets. The audit methodology, the excluded samples and the verified benchmarks are available in the IFM/Arabic-Benchmark-Auditing collection.

Benchmarks marked with ^ will be released by IFM soon.

Quickstart

Serving

vLLM (vllm>=0.30.0) · vLLM recipe:

vllm serve inception42/Jais-3-32B \
  --trust-remote-code \
  --model-impl vllm \
  --dtype bfloat16 \
  --tensor-parallel-size 2 \
  --reasoning-parser k2_horizon \
  --enable-auto-tool-choice \
  --tool-call-parser k2_horizon

Use an exact branch name from the inventory with vLLM's --revision option.

SGLang (sglang>=0.5.20) · SGLang recipe:

sglang serve \
  --model-path inception42/Jais-3-32B \
  --tp 2 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --reasoning-parser k2_horizon \
  --host 0.0.0.0 \
  --port 30000 \
  --tool-call-parser k2_horizon

API Usage

Recommended settings: reasoning_effort="high", the sampling hyperparameters in the table below, and at least 32,768 output tokens. Reasoning depth is selected per request through chat_template_kwargs. Thinking is returned in reasoning_content and the answer in content.

Parameter	Default	Thinking mode, English	Thinking mode, Arabic
temperature	0.2	1.0	0.7
top_p	0.95	0.95	0.95
top_k	20	20	20

Reasoning ExampleTool Calling Example

from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="inception42/Jais-3-32B",
    messages=[{"role": "user", "content": "لدينا عددان مجموعهما 50 وحاصل ضربهما 600. ما هما العددان؟ اشرح خطوات الحل."}],
    temperature=0.7,
    top_p=0.95,
    max_tokens=32768,
    extra_body={"top_k": 20, "chat_template_kwargs": {"reasoning_effort": "high"}},
)
message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)

import json
from datetime import datetime
from zoneinfo import ZoneInfo
from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
TIMEZONES = {"أبوظبي": "Asia/Dubai", "الرياض": "Asia/Riyadh", "القاهرة": "Africa/Cairo", "لندن": "Europe/London"}

def get_current_time(city: str) -> dict:
    if city not in TIMEZONES:
        return {"error": f"مدينة غير معروفة: {city}"}
    return {"city": city, "time": datetime.now(ZoneInfo(TIMEZONES[city])).strftime("%H:%M")}

tools = [{
    "type": "function",
    "function": {
        "name": "get_current_time",
        "description": "يعيد الوقت الحالي في مدينة معينة",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string", "description": "اسم المدينة بالعربية"}},
            "required": ["city"],
        },
    },
}]
available_tools = {"get_current_time": get_current_time}
messages = [{"role": "user", "content": "كم الساعة الآن في القاهرة؟"}]
for _ in range(5):  # call the model, run any requested tools, repeat until it answers
    message = client.chat.completions.create(
        model="inception42/Jais-3-32B",
        messages=messages,
        tools=tools,
        temperature=0.7,
        top_p=0.95,
        max_tokens=32768,
        extra_body={
            "top_k": 20,
            "chat_template_kwargs": {"reasoning_effort": "high", "tool_call_format": "xml"},
        },
    ).choices[0].message
    messages.append(message)
    if not message.tool_calls:
        break
    for call in message.tool_calls:
        tool = available_tools.get(call.function.name)
        arguments = json.loads(call.function.arguments)
        result = tool(**arguments)
        messages.append({"role": "tool", "tool_call_id": call.id,
                         "content": json.dumps(result, ensure_ascii=False)})

print(message.content)

Our model supports multiple tool-call formats, which can be changed with chat_template_kwargs. The supported values are json, xml, and xml_typed. The default is xml. Keep --tool-call-parser k2_horizon enabled to parse the selected format.

Transformers

Validated with Transformers 5.15.0, PyTorch 2.13.0, Safetensors 0.8.0.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "inception42/Jais-3-32B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", dtype="bfloat16", low_cpu_mem_usage=True, trust_remote_code=True
)
inputs = tokenizer("لدينا عددان مجموعهما 50 وحاصل ضربهما 600. ما هما العددان؟ اشرح خطوات الحل.", return_tensors="pt").to(model.device)
inputs.pop("token_type_ids", None)
outputs = model.generate(**inputs, max_new_tokens=32768, temperature=0.7, top_p=0.95, top_k=20, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Release Artifacts

The table below lists the release artifacts for Jais-3-32B.

Artifact Index

Artifact	Link
Model card	Hugging Face
Training logs	W&B
Blog post	Jais 3: Introducing Arabic-centric Agentic AI
Checkpoints	Checkpoint inventory
Technical report	Coming soon
Code repository	Coming soon

Checkpoint Inventory

Model repository: inception42/Jais-3-32B

Branch names below refer to this repository.

Checkpoint	Branch / repository
CPT Checkpoint	CPT
Cooldown Checkpoint	Cooldown
FTPO Checkpoint	FTPO
SFT Checkpoint	SFT

Best Practices

1. Reasoning effort: always high. All reported results use high reasoning effort. Pass {"chat_template_kwargs": {"reasoning_effort": "high"}} on every request. No reasoning ({"reasoning_effort": "low"}) is not recommended except for research on reasoning efforts.
2. Sampling parameters. Use the values in the table above: temperature=0.2, top_p=0.95, top_k=20 by default. In thinking mode, set temperature=1.0 for English and temperature=0.7 for Arabic. Keep min_p=0.0, presence_penalty=0.0 and repetition_penalty=1.0.
3. Serving. Use the validated SGLang recipe above: BF16, TP=2 on a single H100 node, FlashAttention-3.
4. Parsers. Enable the k2_horizon reasoning parser for chat, and add the k2_horizon tool-call parser for agent use. Leave both off for plain completion-style generation.

Intended Use and Limitations

Intended use: The model is designed for Arabic and English assistant and agentic applications, including reasoning and tool-use tasks, as well as for research and experimentation.

Out of scope: The model should not be used in ways that violate applicable laws or regulations. The following uses are also considered out of scope:

• Harmful or malicious use: generating hateful, extremist, discriminatory, or deliberately misleading content, or using the model to support illegal or harmful activities.
• Sensitive information: using the model to generate, expose, or infer personal, confidential, or otherwise sensitive information about individuals or organizations.
• Unsupported languages: applications that depend on reliable performance in languages other than Arabic or English.
• High-stakes decisions: relying on the model to make medical, legal, financial, or safety-critical decisions without appropriate human oversight.

Limitations: The model can produce inaccurate or made-up content and may reflect biases present in its training data. Evaluate it for your use case and add safeguards before deploying it.

Citation

@misc{jais3_2026,
  title  = {Jais 3: Introducing Arabic-centric Agentic AI},
  author = {{IFM Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/jais3/},
}